search

coreylowman / hanabi-mcts

1 stars 0 forks source link

Rollouts... how to determinize? #4

Open coreylowman opened 3 years ago

coreylowman commented 3 years ago

Only extracting the other players hand might not produce super consistent samples...

Both probably need to be extracted at the start of a rollout. Which basically means don't re deferminize every turn?

coreylowman commented 3 years ago

As of 503ed54 rollouts are now down by tracking private info for both players separately. The private info for self (opponents hand) is given to you, and the private info for opponent (your hand) is sampled once at beginning of the rollout. For each players turn they re determinize using the private info available to them up to this point. A player taking a turn modifies the other players private info (e.g. if i play a card, then my opponents private info has changed)

Though this is essentially sampling one instance of a game for a single rollout... does redeterminizing paper sampling multiple possible games during a rollout?